mongez-agent-kit-configuration

Configuration (agentKit in package.json)

Every agent-kit sync invocation reads an optional agentKit block from the project root's package.json. CLI flags (--target, --projects) still override at call time, but the config is the place to pin choices so contributors don't have to remember them.

{
  "scripts": { "postinstall": "agent-kit sync" },
  "agentKit": {
    "targets": ["claude", "cursor"],
    "pick": {
      "@warlock.js/core": true,
      "@my-org/lib": ["only-this-skill"]
    },
    "omit": {
      "@some-vendor/sdk": true,
      "@warlock.js/core": ["add-connector"]
    },
    "monorepo": {
      "projects": ["backend", "frontend"]
    }
  }
}

Malformed sub-fields are silently dropped — a typo in one entry never blocks the whole sync.

targets

Default skill-sync targets when the CLI omits --target. Valid values: claude, copilot, cursor, codex, opencode, amp, goose, kiro, antigravity.

{ "agentKit": { "targets": ["claude", "cursor"] } }

Resolution order: --target flag → agentKit.targets → built-in default (["claude"]). An explicit empty array ("targets": []) is honored but warns — it syncs zero targets, almost always a mistake.

agent-kit init seeds this field for you so you never have to discover the syntax by hand: it writes "targets": ["claude"] (the built-in default, made explicit) when none exists, or whatever you pass to init --target a,b. An already-set targets is left untouched unless you pass --target. See CLI usage.

targets gates only the skills export. The derive step always emits CLAUDE.md, .gemini/GEMINI.md, .github/copilot-instructions.md, and CONVENTIONS.md regardless.

pick — allowlist

When pick is set, only the listed packages are synced; everything else is dropped. Keys are exact package names.

• true → include the whole package's skills.
• string[] → include only the named skills (by their source folder name).